Google has quietly rolled out an update to its open-source AI model, Gemma 4, addressing performance issues and bugs without changing the model's name. The update includes enhancements that improve processing speed on Nvidia Hopper GPUs, with performance gains ranging from 25 to 70 percent. Time to first token has also been reduced by up to 31 percent, according to the company. The update also fixes tool calling bugs, a feature that allows the model to trigger external tools autonomously. Google said the improvements are particularly notable in the telecommunications use case, where agentic reasoning and tool-calling performance increased by 10.1 percent. The update also addresses problems with truncated responses, ensuring more complete outputs for users. Additionally, users can now manually increase the 'max_soft_tokens' parameter for image processing tasks, allowing for sharper OCR results and support for resolutions up to 2.51 megapixels. Google has provided an interactive configurator on Hugging Face for these adjustments.
The published benchmarks only compare the 31B and E4B variants against their predecessors, but the Hugging Face repository shows that all parameter sizes in this model generation have been updated, including the newest 12B release. The community has expressed concerns about Google releasing the update under the same 'Gemma 4' name instead of labeling it as a separate version, such as 'Gemma 4.1.'
Source: thedecoder
